llm - 2024_02
Navigation
Papers
The growing availability of generative AI technologies such as large language models (LLMs) has significant implications for creative work. This paper explores twofold aspects of integrating LLMs into the creative process - the divergence stage of idea generation, and the convergence stage of evaluation and selection of ideas. We devised a collaborative group-AI Brainwriting ideation framework, which incorporated an LLM as an enhancement into the group ideation process, and evaluated the idea generation process and the resulted solution space. To assess the potential of using LLMs in the idea evaluation process, we design an evaluation engine and compared it to idea ratings assigned by three expert and six novice evaluators. Our findings suggest that integrating LLM in Brainwriting could enhance both the ideation process and its outcome. We also provide evidence that LLMs can support idea evaluation. We conclude by discussing implications for HCI education and practice.
Fine-tuning LLMs is crucial to enhancing their task-specific performance and ensuring model behaviors are aligned with human preferences. Among various fine-tuning methods, LoRA is popular for its efficiency and ease to use, allowing end-users to easily post and adopt lightweight LoRA modules on open-source platforms to tailor their model for different customization. However, such a handy share-and-play setting opens up new attack surfaces, that the attacker can render LoRA as an attacker, such as backdoor injection, and widely distribute the adversarial LoRA to the community easily. This can result in detrimental outcomes. Despite the huge potential risks of sharing LoRA modules, this aspect however has not been fully explored. To fill the gap, in this study we thoroughly investigate the attack opportunities enabled in the growing share-and-play scenario. Specifically, we study how to inject backdoor into the LoRA module and dive deeper into LoRA's infection mechanisms. We found that training-free mechanism is possible in LoRA backdoor injection. We also discover the impact of backdoor attacks with the presence of multiple LoRA adaptions concurrently as well as LoRA based backdoor transferability. Our aim is to raise awareness of the potential risks under the emerging share-and-play scenario, so as to proactively prevent potential consequences caused by LoRA-as-an-Attack. Warning: the paper contains potential offensive content generated by models.
We explore the viability of Large Language Models (LLMs), specifically OpenAI's GPT-3.5 and GPT-4, in emulating human survey respondents and eliciting preferences, with a focus on intertemporal choices. Leveraging the extensive literature on intertemporal discounting for benchmarking, we examine responses from LLMs across various languages and compare them to human responses, exploring preferences between smaller, sooner, and larger, later rewards. Our findings reveal that both GPT models demonstrate less patience than humans, with GPT-3.5 exhibiting a lexicographic preference for earlier rewards, unlike human decision-makers. Though GPT-4 does not display lexicographic preferences, its measured discount rates are still considerably larger than those found in humans. Interestingly, GPT models show greater patience in languages with weak future tense references, such as German and Mandarin, aligning with existing literature that suggests a correlation between language structure and intertemporal preferences. We demonstrate how prompting GPT to explain its decisions, a procedure we term "chain-of-thought conjoint," can mitigate, but does not eliminate, discrepancies between LLM and human responses. While directly eliciting preferences using LLMs may yield misleading results, combining chain-of-thought conjoint with topic modeling aids in hypothesis generation, enabling researchers to explore the underpinnings of preferences. Chain-of-thought conjoint provides a structured framework for marketers to use LLMs to identify potential attributes or factors that can explain preference heterogeneity across different customers and contexts.
With the advent of large language models (LLMs) like GPT-3, a natural question is the extent to which these models can be utilized for source code optimization. This paper presents methodologically stringent case studies applied to well-known open source python libraries pillow and numpy. We find that contemporary LLM ChatGPT-4 (state September and October 2023) is surprisingly adept at optimizing energy and compute efficiency. However, this is only the case in interactive use, with a human expert in the loop. Aware of experimenter bias, we document our qualitative approach in detail, and provide transcript and source code. We start by providing a detailed description of our approach in conversing with the LLM to optimize the _getextrema function in the pillow library, and a quantitative evaluation of the performance improvement. To demonstrate qualitative replicability, we report further attempts on another locus in the pillow library, and one code locus in the numpy library, to demonstrate generalization within and beyond a library. In all attempts, the performance improvement is significant (factor up to 38). We have also not omitted reporting of failed attempts (there were none). We conclude that LLMs are a promising tool for code optimization in open source libraries, but that the human expert in the loop is essential for success. Nonetheless, we were surprised by how few iterations were required to achieve substantial performance improvements that were not obvious to the expert in the loop. We would like bring attention to the qualitative nature of this study, more robust quantitative studies would need to introduce a layer of selecting experts in a representative sample -- we invite the community to collaborate.
Large Language Models (LLMs), excel in natural language understanding, but their capability for complex mathematical reasoning with an amalgamation of structured tables and unstructured text is uncertain. This study explores LLMs' mathematical reasoning on four financial tabular question-answering datasets: TATQA, FinQA, ConvFinQA, and Multihiertt. Through extensive experiments with various models and prompting techniques, we assess how LLMs adapt to complex tables and mathematical tasks. We focus on sensitivity to table complexity and performance variations with an increasing number of arithmetic reasoning steps. The results provide insights into LLMs' capabilities and limitations in handling complex mathematical scenarios for semi-structured tables. Ultimately, we introduce a novel prompting technique tailored to semi-structured documents, matching or outperforming other baselines in performance while providing a nuanced understanding of LLMs abilities for such a task.
In this study, we explore the application of Large Language Models (LLMs) in \textit{Jubensha}, a Chinese detective role-playing game and a novel area in Artificial Intelligence (AI) driven gaming. We introduce the first dataset specifically for Jubensha, including character scripts and game rules, to foster AI agent development in this complex narrative environment. Our work also presents a unique multi-agent interaction framework using LLMs, allowing AI agents to autonomously engage in this game. To evaluate the gaming performance of these AI agents, we developed novel methods measuring their mastery of case information and reasoning skills. Furthermore, we incorporated the latest advancements in in-context learning to improve the agents' performance in information gathering, murderer identification, and logical reasoning. The experimental results validate the effectiveness of our proposed methods. This work aims to offer a novel perspective on understanding LLM capabilities and establish a new benchmark for evaluating large language model-based agents.
Large Language Models (LLMs) represent an advanced evolution of earlier, simpler language models. They boast enhanced abilities to handle complex language patterns and generate coherent text, images, audios, and videos. Furthermore, they can be fine-tuned for specific tasks. This versatility has led to the proliferation and extensive use of numerous commercialized large models. However, the rapid expansion of LLMs has raised security and ethical concerns within the academic community. This emphasizes the need for ongoing research into security evaluation during their development and deployment. Over the past few years, a substantial body of research has been dedicated to the security evaluation of large-scale models. This article an in-depth review of the most recent advancements in this field, providing a comprehensive analysis of commonly used evaluation metrics, advanced evaluation frameworks, and the routine evaluation processes for LLMs. Furthermore, we also discuss the future directions for advancing the security evaluation of LLMs.
Effective diabetes management is crucial for maintaining health in diabetic patients. Large Language Models (LLMs) have opened new avenues for diabetes management, facilitating their efficacy. However, current LLM-based approaches are limited by their dependence on general sources and lack of integration with domain-specific knowledge, leading to inaccurate responses. In this paper, we propose a knowledge-infused LLM-powered conversational health agent (CHA) for diabetic patients. We customize and leverage the open-source openCHA framework, enhancing our CHA with external knowledge and analytical capabilities. This integration involves two key components: 1) incorporating the American Diabetes Association dietary guidelines and the Nutritionix information and 2) deploying analytical tools that enable nutritional intake calculation and comparison with the guidelines. We compare the proposed CHA with GPT4. Our evaluation includes 100 diabetes-related questions on daily meal choices and assessing the potential risks associated with the suggested diet. Our findings show that the proposed agent demonstrates superior performance in generating responses to manage essential nutrients.
This paper presents FoFo, a pioneering benchmark for evaluating large language models' (LLMs) ability to follow complex, domain-specific formats, a crucial yet underexamined capability for their application as AI agents. Despite LLMs' advancements, existing benchmarks fail to assess their format-following proficiency adequately. FoFo fills this gap with a diverse range of real-world formats and instructions, developed through an AI-Human collaborative method. Our evaluation across both open-source (e.g., Llama 2, WizardLM) and closed-source (e.g., GPT-4, PALM2, Gemini) LLMs highlights three key findings: open-source models significantly lag behind closed-source ones in format adherence; LLMs' format-following performance is independent of their content generation quality; and LLMs' format proficiency varies across different domains. These insights suggest the need for specialized tuning for format-following skills and highlight FoFo's role in guiding the selection of domain-specific AI agents. FoFo is released here at https://github.com/SalesforceAIResearch/FoFo.
Large Language Model (LLM) systems are inherently compositional, with individual LLM serving as the core foundation with additional layers of objects such as plugins, sandbox, and so on. Along with the great potential, there are also increasing concerns over the security of such probabilistic intelligent systems. However, existing studies on LLM security often focus on individual LLM, but without examining the ecosystem through the lens of LLM systems with other objects (e.g., Frontend, Webtool, Sandbox, and so on). In this paper, we systematically analyze the security of LLM systems, instead of focusing on the individual LLMs. To do so, we build on top of the information flow and formulate the security of LLM systems as constraints on the alignment of the information flow within LLM and between LLM and other objects. Based on this construction and the unique probabilistic nature of LLM, the attack surface of the LLM system can be decomposed into three key components: (1) multi-layer security analysis, (2) analysis of the existence of constraints, and (3) analysis of the robustness of these constraints. To ground this new attack surface, we propose a multi-layer and multi-step approach and apply it to the state-of-art LLM system, OpenAI GPT4. Our investigation exposes several security issues, not just within the LLM model itself but also in its integration with other components. We found that although the OpenAI GPT4 has designed numerous safety constraints to improve its safety features, these safety constraints are still vulnerable to attackers. To further demonstrate the real-world threats of our discovered vulnerabilities, we construct an end-to-end attack where an adversary can illicitly acquire the user's chat history, all without the need to manipulate the user's input or gain direct access to OpenAI GPT4. Our demo is in the link: https://fzwark.github.io/LLM-System-Attack-Demo/
Employing Large Language Models (LLM) in various downstream applications such as classification is crucial, especially for smaller companies lacking the expertise and resources required for fine-tuning a model. Fairness in LLMs helps ensure inclusivity, equal representation based on factors such as race, gender and promotes responsible AI deployment. As the use of LLMs has become increasingly prevalent, it is essential to assess whether LLMs can generate fair outcomes when subjected to considerations of fairness. In this study, we introduce a framework outlining fairness regulations aligned with various fairness definitions, with each definition being modulated by varying degrees of abstraction. We explore the configuration for in-context learning and the procedure for selecting in-context demonstrations using RAG, while incorporating fairness rules into the process. Experiments conducted with different LLMs indicate that GPT-4 delivers superior results in terms of both accuracy and fairness compared to other models. This work is one of the early attempts to achieve fairness in prediction tasks by utilizing LLMs through in-context learning.
Building efficient neural network architectures can be a time-consuming task requiring extensive expert knowledge. This task becomes particularly challenging for edge devices because one has to consider parameters such as power consumption during inferencing, model size, inferencing speed, and CO2 emissions. In this article, we introduce a novel framework designed to automatically discover new neural network architectures based on user-defined parameters, an expert system, and an LLM trained on a large amount of open-domain knowledge. The introduced framework (LeMo-NADe) is tailored to be used by non-AI experts, does not require a predetermined neural architecture search space, and considers a large set of edge device-specific parameters. We implement and validate this proposed neural architecture discovery framework using CIFAR-10, CIFAR-100, and ImageNet16-120 datasets while using GPT-4 Turbo and Gemini as the LLM component. We observe that the proposed framework can rapidly (within hours) discover intricate neural network models that perform extremely well across a diverse set of application settings defined by the user.
The Tulving Test was designed to investigate memory performance in recognition and recall tasks. Its results help assess the relevance of the "Synergistic Ecphory Model" of memory and similar RK paradigms in human performance. This paper starts investigating whether the more than forty-year-old framework sheds some light on LLMs' acts of remembering.
Recent progress in LLMs discussion suggests that multi-agent discussion improves the reasoning abilities of LLMs. In this work, we reevaluate this claim through systematic experiments, where we propose a novel group discussion framework to enrich the set of discussion mechanisms. Interestingly, our results show that a single-agent LLM with strong prompts can achieve almost the same performance as the best existing discussion approach on a wide range of reasoning tasks and backbone LLMs. We observe that the multi-agent discussion performs better than a single agent only when there is no demonstration in the prompt. Further study reveals the common interaction mechanisms of LLMs during the discussion.
Large language models (LLMs) can label data faster and cheaper than humans for various NLP tasks. Despite their prowess, LLMs may fall short in understanding of complex, sociocultural, or domain-specific context, potentially leading to incorrect annotations. Therefore, we advocate a collaborative approach where humans and LLMs work together to produce reliable and high-quality labels. We present MEGAnno+, a human-LLM collaborative annotation system that offers effective LLM agent and annotation management, convenient and robust LLM annotation, and exploratory verification of LLM labels by humans.
This survey provides a comprehensive review of research on multi-turn dialogue systems, with a particular focus on multi-turn dialogue systems based on large language models (LLMs). This paper aims to (a) give a summary of existing LLMs and approaches for adapting LLMs to downstream tasks; (b) elaborate recent advances in multi-turn dialogue systems, covering both LLM-based open-domain dialogue (ODD) and task-oriented dialogue (TOD) systems, along with datasets and evaluation metrics; (c) discuss some future emphasis and recent research problems arising from the development of LLMs and the increasing demands on multi-turn dialogue systems.
We propose ProtLLM, a versatile cross-modal large language model (LLM) for both protein-centric and protein-language tasks. ProtLLM features a unique dynamic protein mounting mechanism, enabling it to handle complex inputs where the natural language text is interspersed with an arbitrary number of proteins. Besides, we propose the protein-as-word language modeling approach to train ProtLLM. By developing a specialized protein vocabulary, we equip the model with the capability to predict not just natural language but also proteins from a vast pool of candidates. Additionally, we construct a large-scale interleaved protein-text dataset, named InterPT, for pre-training. This dataset comprehensively encompasses both (1) structured data sources like protein annotations and (2) unstructured data sources like biological research papers, thereby endowing ProtLLM with crucial knowledge for understanding proteins. We evaluate ProtLLM on classic supervised protein-centric tasks and explore its novel protein-language applications. Experimental results demonstrate that ProtLLM not only achieves superior performance against protein-specialized baselines on protein-centric tasks but also induces zero-shot and in-context learning capabilities on protein-language tasks.
Cognitive behavioral therapy (CBT) is a widely used therapeutic method for guiding individuals toward restructuring their thinking patterns as a means of addressing anxiety, depression, and other challenges. We developed a large language model (LLM)-powered prompt-engineered socially assistive robot (SAR) that guides participants through interactive CBT at-home exercises. We evaluated the performance of the SAR through a 15-day study with 38 university students randomly assigned to interact daily with the robot or a chatbot (using the same LLM), or complete traditional CBT worksheets throughout the duration of the study. We measured weekly therapeutic outcomes, changes in pre-/post-session anxiety measures, and adherence to completing CBT exercises. We found that self-reported measures of general psychological distress significantly decreased over the study period in the robot and worksheet conditions but not the chatbot condition. Furthermore, the SAR enabled significant single-session improvements for more sessions than the other two conditions combined. Our findings suggest that SAR-guided LLM-powered CBT may be as effective as traditional worksheet methods in supporting therapeutic progress from the beginning to the end of the study and superior in decreasing user anxiety immediately after completing the CBT exercise.
Existing question answering (QA) datasets are no longer challenging to most powerful Large Language Models (LLMs). Traditional QA benchmarks like TriviaQA, NaturalQuestions, ELI5 and HotpotQA mainly study ``known unknowns'' with clear indications of both what information is missing, and how to find it to answer the question. Hence, good performance on these benchmarks provides a false sense of security. A yet unmet need of the NLP community is a bank of non-factoid, multi-perspective questions involving a great deal of unclear information needs, i.e. ``unknown uknowns''. We claim we can find such questions in search engine logs, which is surprising because most question-intent queries are indeed factoid. We present Researchy Questions, a dataset of search engine queries tediously filtered to be non-factoid, ``decompositional'' and multi-perspective. We show that users spend a lot of ``effort'' on these questions in terms of signals like clicks and session length, and that they are also challenging for GPT-4. We also show that ``slow thinking'' answering techniques, like decomposition into sub-questions shows benefit over answering directly. We release $\sim$ 100k Researchy Questions, along with the Clueweb22 URLs that were clicked.
Recent research, such as BitNet, is paving the way for a new era of 1-bit Large Language Models (LLMs). In this work, we introduce a 1-bit LLM variant, namely BitNet b1.58, in which every single parameter (or weight) of the LLM is ternary {-1, 0, 1}. It matches the full-precision (i.e., FP16 or BF16) Transformer LLM with the same model size and training tokens in terms of both perplexity and end-task performance, while being significantly more cost-effective in terms of latency, memory, throughput, and energy consumption. More profoundly, the 1.58-bit LLM defines a new scaling law and recipe for training new generations of LLMs that are both high-performance and cost-effective. Furthermore, it enables a new computation paradigm and opens the door for designing specific hardware optimized for 1-bit LLMs.
Existing works on long-term open-domain dialogues focus on evaluating model responses within contexts spanning no more than five chat sessions. Despite advancements in long-context large language models (LLMs) and retrieval augmented generation (RAG) techniques, their efficacy in very long-term dialogues remains unexplored. To address this research gap, we introduce a machine-human pipeline to generate high-quality, very long-term dialogues by leveraging LLM-based agent architectures and grounding their dialogues on personas and temporal event graphs. Moreover, we equip each agent with the capability of sharing and reacting to images. The generated conversations are verified and edited by human annotators for long-range consistency and grounding to the event graphs. Using this pipeline, we collect LoCoMo, a dataset of very long-term conversations, each encompassing 300 turns and 9K tokens on avg., over up to 35 sessions. Based on LoCoMo, we present a comprehensive evaluation benchmark to measure long-term memory in models, encompassing question answering, event summarization, and multi-modal dialogue generation tasks. Our experimental results indicate that LLMs exhibit challenges in understanding lengthy conversations and comprehending long-range temporal and causal dynamics within dialogues. Employing strategies like long-context LLMs or RAG can offer improvements but these models still substantially lag behind human performance.
Large Language Models (LLMs), used in creative writing, code generation, and translation, generate text based on input sequences but are vulnerable to jailbreak attacks, where crafted prompts induce harmful outputs. Most jailbreak prompt methods use a combination of jailbreak templates followed by questions to ask to create jailbreak prompts. However, existing jailbreak prompt designs generally suffer from excessive semantic differences, resulting in an inability to resist defenses that use simple semantic metrics as thresholds. Jailbreak prompts are semantically more varied than the original questions used for queries. In this paper, we introduce a Semantic Mirror Jailbreak (SMJ) approach that bypasses LLMs by generating jailbreak prompts that are semantically similar to the original question. We model the search for jailbreak prompts that satisfy both semantic similarity and jailbreak validity as a multi-objective optimization problem and employ a standardized set of genetic algorithms for generating eligible prompts. Compared to the baseline AutoDAN-GA, SMJ achieves attack success rates (ASR) that are at most 35.4% higher without ONION defense and 85.2% higher with ONION defense. SMJ's better performance in all three semantic meaningfulness metrics of Jailbreak Prompt, Similarity, and Outlier, also means that SMJ is resistant to defenses that use those metrics as thresholds.
Cyberbullying harms teenagers' mental health, and teaching them upstanding intervention is crucial. Wizard-of-Oz studies show chatbots can scale up personalized and interactive cyberbullying education, but implementing such chatbots is a challenging and delicate task. We created a no-code chatbot design tool for K-12 teachers. Using large language models and prompt chaining, our tool allows teachers to prototype bespoke dialogue flows and chatbot utterances. In offering this tool, we explore teachers' distinctive needs when designing chatbots to assist their teaching, and how chatbot design tools might better support them. Our findings reveal that teachers welcome the tool enthusiastically. Moreover, they see themselves as playwrights guiding both the students' and the chatbot's behaviors, while allowing for some improvisation. Their goal is to enable students to rehearse both desirable and undesirable reactions to cyberbullying in a safe environment. We discuss the design opportunities LLM-Chains offer for empowering teachers and the research opportunities this work opens up.
Decision-making is a fundamental capability in everyday life. Large Language Models (LLMs) provide multifaceted support in enhancing human decision-making processes. However, understanding the influencing factors of LLM-assisted decision-making is crucial for enabling individuals to utilize LLM-provided advantages and minimize associated risks in order to make more informed and better decisions. This study presents the results of a comprehensive literature analysis, providing a structural overview and detailed analysis of determinants impacting decision-making with LLM support. In particular, we explore the effects of technological aspects of LLMs, including transparency and prompt engineering, psychological factors such as emotions and decision-making styles, as well as decision-specific determinants such as task difficulty and accountability. In addition, the impact of the determinants on the decision-making process is illustrated via multiple application scenarios. Drawing from our analysis, we develop a dependency framework that systematizes possible interactions in terms of reciprocal interdependencies between these determinants. Our research reveals that, due to the multifaceted interactions with various determinants, factors such as trust in or reliance on LLMs, the user's mental model, and the characteristics of information processing are identified as significant aspects influencing LLM-assisted decision-making processes. Our findings can be seen as crucial for improving decision quality in human-AI collaboration, empowering both users and organizations, and designing more effective LLM interfaces. Additionally, our work provides a foundation for future empirical investigations on the determinants of decision-making assisted by LLMs.
Although the rise of Large Language Models (LLMs) in enterprise settings brings new opportunities and capabilities, it also brings challenges, such as the risk of generating inappropriate, biased, or misleading content that violates regulations and can have legal concerns. To alleviate this, we present "LLMGuard", a tool that monitors user interactions with an LLM application and flags content against specific behaviours or conversation topics. To do this robustly, LLMGuard employs an ensemble of detectors.
Introduced to enhance the efficiency of large language model (LLM) inference, speculative decoding operates by having a smaller model generate a draft. A larger target model then reviews this draft to align with its output, and any acceptance by the target model results in a reduction of the number of the target model runs, ultimately improving efficiency. However, the drafting process in speculative decoding includes slow autoregressive generation and allocates equal time to generating tokens, irrespective of their importance. These inefficiencies collectively contribute to the suboptimal performance of speculative decoding. To further improve LLM inference, we introduce Cascade Speculative Drafting (CS Drafting), a speculative execution algorithm that incorporates two types of cascades. The Vertical Cascade eliminates autoregressive generation from neural models, while the Horizontal Cascade optimizes time allocation in drafting for improved efficiency. Combining both cascades, CS Drafting achieves up to an 81 percent additional speedup over speculative decoding in our experiments, while maintaining the same output distribution as the target model. Our code is publicly available at https://github.com/lfsszd/CS-Drafting.
While large language models (LLMs) often adopt finetuning to unlock their capabilities for downstream applications, our understanding on the inductive biases (especially the scaling properties) of different finetuning methods is still limited. To fill this gap, we conduct systematic experiments studying whether and how different scaling factors, including LLM model size, pretraining data size, new finetuning parameter size and finetuning data size, affect the finetuning performance. We consider two types of finetuning -- full-model tuning (FMT) and parameter efficient tuning (PET, including prompt tuning and LoRA), and explore their scaling behaviors in the data-limited regime where the LLM model size substantially outweighs the finetuning data size. Based on two sets of pretrained bilingual LLMs from 1B to 16B and experiments on bilingual machine translation and multilingual summarization benchmarks, we find that 1) LLM finetuning follows a powerbased multiplicative joint scaling law between finetuning data size and each other scaling factor; 2) LLM finetuning benefits more from LLM model scaling than pretraining data scaling, and PET parameter scaling is generally ineffective; and 3) the optimal finetuning method is highly task- and finetuning data-dependent. We hope our findings could shed light on understanding, selecting and developing LLM finetuning methods.
Semantic Overlap Summarization (SOS) is a constrained multi-document summarization task, where the constraint is to capture the common/overlapping information between two alternative narratives. While recent advancements in Large Language Models (LLMs) have achieved superior performance in numerous summarization tasks, a benchmarking study of the SOS task using LLMs is yet to be performed. As LLMs' responses are sensitive to slight variations in prompt design, a major challenge in conducting such a benchmarking study is to systematically explore a variety of prompts before drawing a reliable conclusion. Fortunately, very recently, the TELeR taxonomy has been proposed which can be used to design and explore various prompts for LLMs. Using this TELeR taxonomy and 15 popular LLMs, this paper comprehensively evaluates LLMs on the SOS Task, assessing their ability to summarize overlapping information from multiple alternative narratives. For evaluation, we report well-established metrics like ROUGE, BERTscore, and SEM-F1$ on two different datasets of alternative narratives. We conclude the paper by analyzing the strengths and limitations of various LLMs in terms of their capabilities in capturing overlapping information The code and datasets used to conduct this study are available at https://anonymous.4open.science/r/llm_eval-E16D.
With the fast development of large language models (LLMs), LLM-driven Web Agents (Web Agents for short) have obtained tons of attention due to their superior capability where LLMs serve as the core part of making decisions like the human brain equipped with multiple web tools to actively interact with external deployed websites. As uncountable Web Agents have been released and such LLM systems are experiencing rapid development and drawing closer to widespread deployment in our daily lives, an essential and pressing question arises: "Are these Web Agents secure?". In this paper, we introduce a novel threat, WIPI, that indirectly controls Web Agent to execute malicious instructions embedded in publicly accessible webpages. To launch a successful WIPI works in a black-box environment. This methodology focuses on the form and content of indirect instructions within external webpages, enhancing the efficiency and stealthiness of the attack. To evaluate the effectiveness of the proposed methodology, we conducted extensive experiments using 7 plugin-based ChatGPT Web Agents, 8 Web GPTs, and 3 different open-source Web Agents. The results reveal that our methodology achieves an average attack success rate (ASR) exceeding 90% even in pure black-box scenarios. Moreover, through an ablation study examining various user prefix instructions, we demonstrated that the WIPI exhibits strong robustness, maintaining high performance across diverse prefix instructions.
Machine unlearning can be useful for removing harmful capabilities and memorized text from large language models (LLMs), but there are not yet standardized methods for rigorously evaluating it. In this paper, we first survey techniques and limitations of existing unlearning evaluations. Second, we apply a comprehensive set of tests for the robustness and competitiveness of unlearning in the "Who's Harry Potter" (WHP) model from Eldan and Russinovich (2023). While WHP's unlearning generalizes well when evaluated with the "Familiarity" metric from Eldan and Russinovich, we find i) higher-than-baseline amounts of knowledge can reliably be extracted, ii) WHP performs on par with the original model on Harry Potter Q&A tasks, iii) it represents latent knowledge comparably to the original model, and iv) there is collateral unlearning in related domains. Overall, our results highlight the importance of comprehensive unlearning evaluation that avoids ad-hoc metrics.
AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. Proximal Policy Optimization (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit the formulation of alignment from human preferences in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed "RL-free" methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics enables benefiting from online RL optimization at low cost.
Evaluating natural language generation (NLG) is a vital but challenging problem in artificial intelligence. Traditional evaluation metrics mainly capturing content (e.g. n-gram) overlap between system outputs and references are far from satisfactory, and large language models (LLMs) such as ChatGPT have demonstrated great potential in NLG evaluation in recent years. Various automatic evaluation methods based on LLMs have been proposed, including metrics derived from LLMs, prompting LLMs, and fine-tuning LLMs with labeled evaluation data. In this survey, we first give a taxonomy of LLM-based NLG evaluation methods, and discuss their pros and cons, respectively. We also discuss human-LLM collaboration for NLG evaluation. Lastly, we discuss several open problems in this area and point out future research directions.
The recent development and success of Large Language Models (LLMs) necessitate an evaluation of their performance across diverse NLP tasks in different languages. Although several frameworks have been developed and made publicly available, their customization capabilities for specific tasks and datasets are often complex for different users. In this study, we introduce the LLMeBench framework, which can be seamlessly customized to evaluate LLMs for any NLP task, regardless of language. The framework features generic dataset loaders, several model providers, and pre-implements most standard evaluation metrics. It supports in-context learning with zero- and few-shot settings. A specific dataset and task can be evaluated for a given LLM in less than 20 lines of code while allowing full flexibility to extend the framework for custom datasets, models, or tasks. The framework has been tested on 31 unique NLP tasks using 53 publicly available datasets within 90 experimental setups, involving approximately 296K data points. We open-sourced LLMeBench for the community (https://github.com/qcri/LLMeBench/) and a video demonstrating the framework is available online. (https://youtu.be/9cC2m_abk3A)
We announce the initial release of "Airavata," an instruction-tuned LLM for Hindi. Airavata was created by fine-tuning OpenHathi with diverse, instruction-tuning Hindi datasets to make it better suited for assistive tasks. Along with the model, we also share the IndicInstruct dataset, which is a collection of diverse instruction-tuning datasets to enable further research for Indic LLMs. Additionally, we present evaluation benchmarks and a framework for assessing LLM performance across tasks in Hindi. Currently, Airavata supports Hindi, but we plan to expand this to all 22 scheduled Indic languages. You can access all artifacts at https://ai4bharat.github.io/airavata.
As sufficient data are not always publically accessible for model training, researchers exploit limited data with advanced learning algorithms or expand the dataset via data augmentation (DA). Conducting DA in private domain requires private protection approaches (i.e. anonymization and perturbation), but those methods cannot provide protection guarantees. Differential privacy (DP) learning methods theoretically bound the protection but are not skilled at generating pseudo text samples with large models. In this paper, we transfer DP-based pseudo sample generation task to DP-based generated samples discrimination task, where we propose a DP-based DA method with a LLM and a DP-based discriminator for text classification on private domains. We construct a knowledge distillation model as the DP-based discriminator: teacher models, accessing private data, teaches students how to select private samples with calibrated noise to achieve DP. To constrain the distribution of DA's generation, we propose a DP-based tutor that models the noised private distribution and controls samples' generation with a low privacy cost. We theoretically analyze our model's privacy protection and empirically verify our model.
We aim to present a comprehensive overview of the latest advancements in utilizing Large Language Models (LLMs) within the healthcare sector, emphasizing their transformative impact across various medical domains. LLMs have become pivotal in supporting healthcare, including physicians, healthcare providers, and patients. Our review provides insight into the applications of Large Language Models (LLMs) in healthcare, specifically focusing on diagnostic and treatment-related functionalities. We shed light on how LLMs are applied in cancer care, dermatology, dental care, neurodegenerative disorders, and mental health, highlighting their innovative contributions to medical diagnostics and patient care. Throughout our analysis, we explore the challenges and opportunities associated with integrating LLMs in healthcare, recognizing their potential across various medical specialties despite existing limitations. Additionally, we offer an overview of handling diverse data types within the medical field.
The ever-increasing large language models (LLMs), though opening a potential path for the upcoming artificial general intelligence, sadly drops a daunting obstacle on the way towards their on-device deployment. As one of the most well-established pre-LLMs approaches in reducing model complexity, network pruning appears to lag behind in the era of LLMs, due mostly to its costly fine-tuning (or re-training) necessity under the massive volumes of model parameter and training data. To close this industry-academia gap, we introduce Dynamic Sparse No Training (DSnoT), a training-free fine-tuning approach that slightly updates sparse LLMs without the expensive backpropagation and any weight updates. Inspired by the Dynamic Sparse Training, DSnoT minimizes the reconstruction error between the dense and sparse LLMs, in the fashion of performing iterative weight pruning-and-growing on top of sparse LLMs. To accomplish this purpose, DSnoT particularly takes into account the anticipated reduction in reconstruction error for pruning and growing, as well as the variance w.r.t. different input data for growing each weight. This practice can be executed efficiently in linear time since its obviates the need of backpropagation for fine-tuning LLMs. Extensive experiments on LLaMA-V1/V2, Vicuna, and OPT across various benchmarks demonstrate the effectiveness of DSnoT in enhancing the performance of sparse LLMs, especially at high sparsity levels. For instance, DSnoT is able to outperform the state-of-the-art Wanda by 26.79 perplexity at 70% sparsity with LLaMA-7B. Our paper offers fresh insights into how to fine-tune sparse LLMs in an efficient training-free manner and open new venues to scale the great potential of sparsity to LLMs. Codes are available at https://github.com/zyxxmu/DSnoT.
Diffusion-based generative models have significantly advanced text-to-image generation but encounter challenges when processing lengthy and intricate text prompts describing complex scenes with multiple objects. While excelling in generating images from short, single-object descriptions, these models often struggle to faithfully capture all the nuanced details within longer and more elaborate textual inputs. In response, we present a novel approach leveraging Large Language Models (LLMs) to extract critical components from text prompts, including bounding box coordinates for foreground objects, detailed textual descriptions for individual objects, and a succinct background context. These components form the foundation of our layout-to-image generation model, which operates in two phases. The initial Global Scene Generation utilizes object layouts and background context to create an initial scene but often falls short in faithfully representing object characteristics as specified in the prompts. To address this limitation, we introduce an Iterative Refinement Scheme that iteratively evaluates and refines box-level content to align them with their textual descriptions, recomposing objects as needed to ensure consistency. Our evaluation on complex prompts featuring multiple objects demonstrates a substantial improvement in recall compared to baseline diffusion models. This is further validated by a user study, underscoring the efficacy of our approach in generating coherent and detailed scenes from intricate textual inputs.
Timely, personalized feedback is essential for students learning programming. LLM-powered tools like ChatGPT offer instant support, but reveal direct answers with code, which may hinder deep conceptual engagement. We developed CodeAid, an LLM-powered programming assistant delivering helpful, technically correct responses, without revealing code solutions. CodeAid answers conceptual questions, generates pseudo-code with line-by-line explanations, and annotates student's incorrect code with fix suggestions. We deployed CodeAid in a programming class of 700 students for a 12-week semester. A thematic analysis of 8,000 usages of CodeAid was performed, further enriched by weekly surveys, and 22 student interviews. We then interviewed eight programming educators to gain further insights. Our findings reveal four design considerations for future educational AI assistants: D1) exploiting AI's unique benefits; D2) simplifying query formulation while promoting cognitive engagement; D3) avoiding direct responses while encouraging motivated learning; and D4) maintaining transparency and control for students to asses and steer AI responses.
Hallucinations pose a significant challenge to the reliability and alignment of Large Language Models (LLMs), limiting their widespread acceptance beyond chatbot applications. Despite ongoing efforts, hallucinations remain a prevalent challenge in LLMs. The detection of hallucinations itself is also a formidable task, frequently requiring manual labeling or constrained evaluations. This paper introduces an automated scalable framework that combines benchmarking LLMs' hallucination tendencies with efficient hallucination detection. We leverage LLMs to generate challenging tasks related to hypothetical phenomena, subsequently employing them as agents for efficient hallucination detection. The framework is domain-agnostic, allowing the use of any language model for benchmark creation or evaluation in any domain. We introduce the publicly available HypoTermQA Benchmarking Dataset, on which state-of-the-art models' performance ranged between 3% and 11%, and evaluator agents demonstrated a 6% error rate in hallucination prediction. The proposed framework provides opportunities to test and improve LLMs. Additionally, it has the potential to generate benchmarking datasets tailored to specific domains, such as law, health, and finance.
Reinforcement learning (RL) has become the de facto standard practice for sequential decision-making problems by improving future acting policies with feedback. However, RL algorithms may require extensive trial-and-error interactions to collect useful feedback for improvement. On the other hand, recent developments in large language models (LLMs) have showcased impressive capabilities in language understanding and generation, yet they fall short in exploration and self-improvement capabilities for planning tasks, lacking the ability to autonomously refine their responses based on feedback. Therefore, in this paper, we study how the policy prior provided by the LLM can enhance the sample efficiency of RL algorithms. Specifically, we develop an algorithm named LINVIT that incorporates LLM guidance as a regularization factor in value-based RL, leading to significant reductions in the amount of data needed for learning, particularly when the difference between the ideal policy and the LLM-informed policy is small, which suggests that the initial policy is close to optimal, reducing the need for further exploration. Additionally, we present a practical algorithm SLINVIT that simplifies the construction of the value function and employs subgoals to reduce the search complexity. Our experiments across three interactive environments ALFWorld, InterCode, and BlocksWorld demonstrate that our method achieves state-of-the-art success rates and also surpasses previous RL and LLM approaches in terms of sample efficiency. Our code is available at https://github.com/agentification/Language-Integrated-VI.
While Large Language Models (LLMs) demonstrate impressive capabilities in text generation, we find that their ability has yet to be generalized to music, humanity's creative language. We introduce ChatMusician, an open-source LLM that integrates intrinsic musical abilities. It is based on continual pre-training and finetuning LLaMA2 on a text-compatible music representation, ABC notation, and the music is treated as a second language. ChatMusician can understand and generate music with a pure text tokenizer without any external multi-modal neural structures or tokenizers. Interestingly, endowing musical abilities does not harm language abilities, even achieving a slightly higher MMLU score. Our model is capable of composing well-structured, full-length music, conditioned on texts, chords, melodies, motifs, musical forms, etc, surpassing GPT-4 baseline. On our meticulously curated college-level music understanding benchmark, MusicTheoryBench, ChatMusician surpasses LLaMA2 and GPT-3.5 on zero-shot setting by a noticeable margin. Our work reveals that LLMs can be an excellent compressor for music, but there remains significant territory to be conquered. We release our 4B token music-language corpora MusicPile, the collected MusicTheoryBench, code, model and demo in GitHub.
While fine-tuning large language models (LLMs) for specific tasks often yields impressive results, it comes at the cost of memory inefficiency due to back-propagation in gradient-based training. Memory-efficient Zeroth-order (MeZO) optimizers, recently proposed to address this issue, only require forward passes during training, making them more memory-friendly. However, the quality of gradient estimates in zeroth order optimization often depends on the data dimensionality, potentially explaining why MeZO still exhibits significant performance drops compared to standard fine-tuning across various tasks. Inspired by the success of Parameter-Efficient Fine-Tuning (PEFT), this paper introduces Sparse MeZO, a novel memory-efficient zeroth-order optimization approach that applies ZO only to a carefully chosen subset of parameters. We propose a simple yet effective parameter selection scheme that yields significant performance gains with Sparse-MeZO. Additionally, we develop a memory-optimized implementation for sparse masking, ensuring the algorithm requires only inference-level memory consumption, allowing Sparse-MeZO to fine-tune LLaMA-30b on a single A100 GPU. Experimental results illustrate that Sparse-MeZO consistently improves both performance and convergence speed over MeZO without any overhead. For example, it achieves a 9\% absolute accuracy improvement and 3.5x speedup over MeZO on the RTE task.
In recent years, there have been remarkable advancements in node classification achieved by Graph Neural Networks (GNNs). However, they necessitate abundant high-quality labels to ensure promising performance. In contrast, Large Language Models (LLMs) exhibit impressive zero-shot proficiency on text-attributed graphs. Yet, they face challenges in efficiently processing structural data and suffer from high inference costs. In light of these observations, this work introduces a label-free node classification on graphs with LLMs pipeline, LLM-GNN. It amalgamates the strengths of both GNNs and LLMs while mitigating their limitations. Specifically, LLMs are leveraged to annotate a small portion of nodes and then GNNs are trained on LLMs' annotations to make predictions for the remaining large portion of nodes. The implementation of LLM-GNN faces a unique challenge: how can we actively select nodes for LLMs to annotate and consequently enhance the GNN training? How can we leverage LLMs to obtain annotations of high quality, representativeness, and diversity, thereby enhancing GNN performance with less cost? To tackle this challenge, we develop an annotation quality heuristic and leverage the confidence scores derived from LLMs to advanced node selection. Comprehensive experimental results validate the effectiveness of LLM-GNN. In particular, LLM-GNN can achieve an accuracy of 74.9% on a vast-scale dataset \products with a cost less than 1 dollar.
Large language models (LLMs) are documented to struggle in settings that require complex reasoning. Nevertheless, instructing the model to break down the problem into smaller reasoning steps, or ensembling various generations through modifying decoding steps boosts performance. However, these methods assume that the input prompt is fixed and expect the decoding strategies to introduce the diversity needed for ensembling. In this work, we discuss how one can create and leverage variations of the input prompt as a means of diversity of thought. We propose a method that automatically improves prompt diversity by soliciting feedback from the LLM to ideate approaches that are apt for the problem. We then ensemble the diverse prompts in our method DIVSE (DIVerse reasoning path Self-Ensemble) across multiple inference calls, or use diverse approaches within a single inference call; we call the latter IDIV-SE (In-call DIVerse reasoning path Self-Ensemble). Apart from our approaches outperforming prior work, DIV-SE(in particular) advances state-of-the-art performance on the challenging planning and graph coloring benchmarks. Our results improve the Pareto frontier of the accuracy-cost trade-off.
One of the most important yet onerous tasks in the academic peer-reviewing process is composing meta-reviews, which involves understanding the core contributions, strengths, and weaknesses of a scholarly manuscript based on peer-review narratives from multiple experts and then summarizing those multiple experts' perspectives into a concise holistic overview. Given the latest major developments in generative AI, especially Large Language Models (LLMs), it is very compelling to rigorously study the utility of LLMs in generating such meta-reviews in an academic peer-review setting. In this paper, we perform a case study with three popular LLMs, i.e., GPT-3.5, LLaMA2, and PaLM2, to automatically generate meta-reviews by prompting them with different types/levels of prompts based on the recently proposed TELeR taxonomy. Finally, we perform a detailed qualitative study of the meta-reviews generated by the LLMs and summarize our findings and recommendations for prompting LLMs for this complex task.
Large language models (LLMs) are displaying emergent abilities for math reasoning tasks,and there is a growing attention on enhancing the ability of open-source LLMs through supervised fine-tuning (SFT).In this paper, we aim to explore a general data strategy for supervised data to help optimize and expand math reasoning ability.Firstly, we determine the ability boundary of reasoning paths augmentation by identifying these paths' minimal optimal set.Secondly, we validate that different abilities of the model can be cumulatively enhanced by Mix of Minimal Optimal Sets of corresponding types of data, while our models MMOS achieve SOTA performance on series base models under much lower construction costs.Besides, we point out GSM-HARD is not really hard and today's LLMs no longer lack numerical robustness.Also, we provide an Auto Problem Generator for robustness testing and educational applications.Our code and data are publicly available at https://github.com/cyzhh/MMOS.
Large Language Models (LLMs) have shown impressive abilities in data annotation, opening the way for new approaches to solve classic NLP problems. In this paper, we show how to use LLMs to create NuNER, a compact language representation model specialized in the Named Entity Recognition (NER) task. NuNER can be fine-tuned to solve downstream NER problems in a data-efficient way, outperforming similar-sized foundation models in the few-shot regime and competing with much larger LLMs. We find that the size and entity-type diversity of the pre-training dataset are key to achieving good performance. We view NuNER as a member of the broader family of task-specific foundation models, recently unlocked by LLMs.
In this work we show that the size versus accuracy trade-off of neural network quantization can be significantly improved by increasing the quantization dimensionality. We propose the GPTVQ method, a new fast method for post-training vector quantization (VQ) that scales well to Large Language Models (LLMs). Our method interleaves quantization of one or more columns with updates to the remaining unquantized weights, using information from the Hessian of the per-layer output reconstruction MSE. Quantization codebooks are initialized using an efficient data-aware version of the EM algorithm. The codebooks are then updated, and further compressed by using integer quantization and SVD-based compression. GPTVQ establishes a new state-of-the art in the size vs accuracy trade-offs on a wide range of LLMs such as Llama-v2 and Mistral. Furthermore, our method is efficient: on a single H100 it takes between 3 and 11 hours to process a Llamav2-70B model, depending on quantization setting. Lastly, with on-device timings for VQ decompression on a mobile CPU we show that VQ leads to improved latency compared to using a 4-bit integer format.
Extensive studies have been devoted to privatizing general-domain Large Language Models (LLMs) as Domain-Specific LLMs via feeding specific-domain data. However, these privatization efforts often ignored a critical aspect: Dual Logic Ability, which is a core reasoning ability for LLMs. The dual logic ability of LLMs ensures that they can maintain a consistent stance when confronted with both positive and negative statements about the same fact. Our study focuses on how the dual logic ability of LLMs is affected during the privatization process in the medical domain. We conduct several experiments to analyze the dual logic ability of LLMs by examining the consistency of the stance in responses to paired questions about the same fact. In our experiments, interestingly, we observed a significant decrease in the dual logic ability of existing LLMs after privatization. Besides, our results indicate that incorporating general domain dual logic data into LLMs not only enhances LLMs' dual logic ability but also further improves their accuracy. These findings underscore the importance of prioritizing LLMs' dual logic ability during the privatization process. Our study establishes a benchmark for future research aimed at exploring LLMs' dual logic ability during the privatization process and offers valuable guidance for privatization efforts in real-world applications.
Intensive care unit (ICU) patients often develop new health-related problems in their long-term recovery. Health care professionals keeping a diary of a patient's stay is a proven strategy to tackle this but faces several adoption barriers, such as lack of time and difficulty in knowing what to write. Large language models (LLMs), with their ability to generate human-like text and adaptability, could solve these challenges. However, realizing this vision involves addressing several socio-technical and practical research challenges. This paper discusses these challenges and proposes future research directions to utilize the potential of LLMs in ICU diary writing, ultimately improving the long-term recovery outcomes for ICU patients.
Despite making significant progress in multi-modal tasks, current Multi-modal Large Language Models (MLLMs) encounter the significant challenge of hallucinations, which may lead to harmful consequences. Therefore, evaluating MLLMs' hallucinations is becoming increasingly important in model improvement and practical application deployment. Previous works are limited in high evaluation costs (e.g., relying on humans or advanced LLMs) and insufficient evaluation dimensions (e.g., types of tasks and hallucinations). In this paper, we propose an LLM-free multi-dimensional benchmark AMBER, which can be used to evaluate both generative task and discriminative task including existence, attribute and relation hallucination. Based on AMBER, we design a low-cost and efficient evaluation pipeline. Additionally, we conduct a comprehensive evaluation and detailed analysis of mainstream MLLMs including GPT-4V(ision), and also give guideline suggestions for mitigating hallucinations. The data and code of AMBER are available at https://github.com/junyangwang0410/AMBER.
Extensive work has been devoted to improving the safety mechanism of Large Language Models (LLMs). However, LLMs still tend to generate harmful responses when faced with malicious instructions, a phenomenon referred to as "Jailbreak Attack". In our research, we introduce a novel automatic jailbreak method RADIAL, which bypasses the security mechanism by amplifying the potential of LLMs to generate affirmation responses. The jailbreak idea of our method is "Inherent Response Tendency Analysis" which identifies real-world instructions that can inherently induce LLMs to generate affirmation responses and the corresponding jailbreak strategy is "Real-World Instructions-Driven Jailbreak" which involves strategically splicing real-world instructions identified through the above analysis around the malicious instruction. Our method achieves excellent attack performance on English malicious instructions with five open-source advanced LLMs while maintaining robust attack performance in executing cross-language attacks against Chinese malicious instructions. We conduct experiments to verify the effectiveness of our jailbreak idea and the rationality of our jailbreak strategy design. Notably, our method designed a semantically coherent attack prompt, highlighting the potential risks of LLMs. Our study provides detailed insights into jailbreak attacks, establishing a foundation for the development of safer LLMs.
Code editing is an essential step towards reliable program synthesis to automatically correct critical errors generated from code LLMs. Recent studies have demonstrated that closed-source LLMs (i.e., ChatGPT and GPT-4) are capable of generating corrective feedback to edit erroneous inputs. However, it remains challenging for open-source code LLMs to generate feedback for code editing, since these models tend to adhere to the superficial formats of feedback and provide feedback with misleading information. Hence, the focus of our work is to leverage open-source code LLMs to generate helpful feedback with correct guidance for code editing. To this end, we present Coffee, a collected dataset specifically designed for code fixing with feedback. Using this dataset, we construct CoffeePots, a framework for COde Fixing with FEEdback via Preference-Optimized Tuning and Selection. The proposed framework aims to automatically generate helpful feedback for code editing while minimizing the potential risk of superficial feedback. The combination of Coffee and CoffeePots marks a significant advancement, achieving state-of-the-art performance on HumanEvalFix benchmark. Codes and model checkpoints are publicly available at https://github.com/Lune-Blue/COFFEE.
The booming success of LLMs initiates rapid development in LLM agents. Though the foundation of an LLM agent is the generative model, it is critical to devise the optimal reasoning strategies and agent architectures. Accordingly, LLM agent research advances from the simple chain-of-thought prompting to more complex ReAct and Reflection reasoning strategy; agent architecture also evolves from single agent generation to multi-agent conversation, as well as multi-LLM multi-agent group chat. However, with the existing intricate frameworks and libraries, creating and evaluating new reasoning strategies and agent architectures has become a complex challenge, which hinders research investigation into LLM agents. Thus, we open-source a new AI agent library, AgentLite, which simplifies this process by offering a lightweight, user-friendly platform for innovating LLM agent reasoning, architectures, and applications with ease. AgentLite is a task-oriented framework designed to enhance the ability of agents to break down tasks and facilitate the development of multi-agent systems. Furthermore, we introduce multiple practical applications developed with AgentLite to demonstrate its convenience and flexibility. Get started now at: \url{https://github.com/SalesforceAIResearch/AgentLite}.
Large Language models (LLMs) are finding wide use in software engineering practice. These models are extremely data-hungry, and are largely trained on open-source (OSS) code distributed with permissive licenses. In terms of actual use however, a great deal of software development still occurs in the for-profit/proprietary sphere, where the code under development is not, and never has been, in the public domain; thus, many developers, do their work, and use LLMs, in settings where the models may not be as familiar with the code under development. In such settings, do LLMs work as well as they do for OSS code? If not, what are the differences? When performance differs, what are the possible causes, and are there work-arounds? In this paper, we examine this issue using proprietary, closed-source software data from Microsoft, where most proprietary code is in C# and C++. We find that performance for C# changes little from OSS --> proprietary code, but does significantly reduce for C++; we find that this difference is attributable to differences in identifiers. We also find that some performance degradation, in some cases, can be ameliorated efficiently by in-context learning.
The rapid development in the field of Large Language Models (LLMs) has led to a surge in applications that facilitate collaboration among multiple agents to assist humans in their daily tasks. However, a significant gap remains in assessing whether LLM-powered applications genuinely enhance user experience and task execution efficiency. This highlights the pressing need for methods to verify utility of LLM-powered applications, particularly by ensuring alignment between the application's functionality and end-user needs. We introduce AgentEval provides an implementation for the math problems, a novel framework designed to simplify the utility verification process by automatically proposing a set of criteria tailored to the unique purpose of any given application. This allows for a comprehensive assessment, quantifying the utility of an application against the suggested criteria. We present a comprehensive analysis of the robustness of quantifier's work.
Tokenization, the division of input text into input tokens, is an often overlooked aspect of the large language model (LLM) pipeline and could be the source of useful or harmful inductive biases. Historically, LLMs have relied on byte pair encoding, without care to specific input domains. With the increased use of LLMs for reasoning, various number-specific tokenization schemes have been adopted, with popular models like LLaMa and PaLM opting for single-digit tokenization while GPT-3.5 and GPT-4 have separate tokens for each 1-, 2-, and 3-digit numbers. In this work, we study the effect this choice has on numerical reasoning through the use of arithmetic tasks. We consider left-to-right and right-to-left tokenization for GPT-3.5 and -4, finding that right-to-left tokenization (enforced by comma separating numbers at inference time) leads to largely improved performance. Furthermore, we find that model errors when using standard left-to-right tokenization follow stereotyped error patterns, suggesting that model computations are systematic rather than approximate. We show that the model is able to convert between tokenizations easily, thus allowing chain-of-thought-inspired approaches to recover performance on left-to-right tokenized inputs. We also find the gap between tokenization directions decreases when models are scaled, possibly indicating that larger models are better able to override this tokenization-dependent inductive bias. In summary, our work performs the first study of how number tokenization choices lead to differences in model performance on arithmetic tasks, accompanied by a thorough analysis of error patterns. We hope this work inspires practitioners to more carefully ablate number tokenization-related choices when working towards general models of numerical reasoning.
Recent works show that the largest of the large language models (LLMs) can solve many simple reasoning tasks expressed in natural language, without any/much supervision. But, can they also solve challenging first-order combinatorial reasoning problems, such as graph coloring, knapsack and cryptarithmetic? To answer this question, we present PuzzleBench, a dataset of 31 such challenging problems along with a few solved instances for each problem. These problems are all first order, i.e., they can be instantiated with problem instances of varying sizes, and most of them are NP-hard, requiring several reasoning steps to reach the solution. We first observe that LLMs, even when aided by symbolic solvers, perform rather poorly on our dataset. In response, we propose a new approach, Puzzle-LM, which combines LLMs with both symbolic solvers and program interpreters, along with feedback from solved examples, to achieve huge performance gains. Our extensive experimentation and analyses offer new insights into the reasoning abilities and limitations of present-day LLMs.
Large language models (LLMs) have become the secret ingredient driving numerous industrial applications, showcasing their remarkable versatility across a diverse spectrum of tasks. From natural language processing and sentiment analysis to content generation and personalized recommendations, their unparalleled adaptability has facilitated widespread adoption across industries. This transformative shift driven by LLMs underscores the need to explore the underlying associated challenges and avenues for enhancement in their utilization. In this paper, our objective is to unravel and evaluate the obstacles and opportunities inherent in leveraging LLMs within an industrial context. To this end, we conduct a survey involving a group of industry practitioners, develop four research questions derived from the insights gathered, and examine 68 industry papers to address these questions and derive meaningful conclusions.
Large Language Models (LLMs) are capable of generating text that is similar to or surpasses human quality. However, it is unclear whether LLMs tend to exhibit distinctive linguistic styles akin to how human authors do. Through a comprehensive linguistic analysis, we compare the vocabulary, Part-Of-Speech (POS) distribution, dependency distribution, and sentiment of texts generated by three of the most popular LLMS today (GPT-3.5, GPT-4, and Bard) to diverse inputs. The results point to significant linguistic variations which, in turn, enable us to attribute a given text to its LLM origin with a favorable 88\% accuracy using a simple off-the-shelf classification model. Theoretical and practical implications of this intriguing finding are discussed.
Natural Language Processing (NLP) research is increasingly focusing on the use of Large Language Models (LLMs), with some of the most popular ones being either fully or partially closed-source. The lack of access to model details, especially regarding training data, has repeatedly raised concerns about data contamination among researchers. Several attempts have been made to address this issue, but they are limited to anecdotal evidence and trial and error. Additionally, they overlook the problem of \emph{indirect} data leaking, where models are iteratively improved by using data coming from users. In this work, we conduct the first systematic analysis of work using OpenAI's GPT-3.5 and GPT-4, the most prominently used LLMs today, in the context of data contamination. By analysing 255 papers and considering OpenAI's data usage policy, we extensively document the amount of data leaked to these models during the first year after the model's release. We report that these models have been globally exposed to $\sim$4.7M samples from 263 benchmarks. At the same time, we document a number of evaluation malpractices emerging in the reviewed papers, such as unfair or missing baseline comparisons and reproducibility issues. We release our results as a collaborative project on https://leak-llm.github.io/, where other researchers can contribute to our efforts.
MeTMaP: Metamorphic Testing for Detecting False Vector Matching Problems in LLM Augmented Generation
Augmented generation techniques such as Retrieval-Augmented Generation (RAG) and Cache-Augmented Generation (CAG) have revolutionized the field by enhancing large language model (LLM) outputs with external knowledge and cached information. However, the integration of vector databases, which serve as a backbone for these augmentations, introduces critical challenges, particularly in ensuring accurate vector matching. False vector matching in these databases can significantly compromise the integrity and reliability of LLM outputs, leading to misinformation or erroneous responses. Despite the crucial impact of these issues, there is a notable research gap in methods to effectively detect and address false vector matches in LLM-augmented generation. This paper presents MeTMaP, a metamorphic testing framework developed to identify false vector matching in LLM-augmented generation systems. We derive eight metamorphic relations (MRs) from six NLP datasets, which form our method's core, based on the idea that semantically similar texts should match and dissimilar ones should not. MeTMaP uses these MRs to create sentence triplets for testing, simulating real-world LLM scenarios. Our evaluation of MeTMaP over 203 vector matching configurations, involving 29 embedding models and 7 distance metrics, uncovers significant inaccuracies. The results, showing a maximum accuracy of only 41.51\% on our tests compared to the original datasets, emphasize the widespread issue of false matches in vector matching methods and the critical need for effective detection and mitigation in LLM-augmented applications.
Recent years have seen important advances in the building of interpretable models, machine learning models that are designed to be easily understood by humans. In this work, we show that large language models (LLMs) are remarkably good at working with interpretable models, too. In particular, we show that LLMs can describe, interpret, and debug Generalized Additive Models (GAMs). Combining the flexibility of LLMs with the breadth of statistical patterns accurately described by GAMs enables dataset summarization, question answering, and model critique. LLMs can also improve the interaction between domain experts and interpretable models, and generate hypotheses about the underlying phenomenon. We release \url{https://github.com/interpretml/TalkToEBM} as an open-source LLM-GAM interface.
Education that suits the individual learning level is necessary to improve students' understanding. The first step in achieving this purpose by using large language models (LLMs) is to adjust the textual difficulty of the response to students. This work analyzes how LLMs can implicitly adjust text difficulty between user input and its generated text. To conduct the experiments, we created a new dataset from Stack-Overflow to explore the performance of question-answering-based conversation. Experimental results on the Stack-Overflow dataset and the TSCC dataset, including multi-turn conversation show that LLMs can implicitly handle text difficulty between user input and its generated response. We also observed that some LLMs can surpass humans in handling text difficulty and the importance of instruction-tuning.
While enabling large language models to implement function calling (known as APIs) can greatly enhance the performance of Large Language Models (LLMs), function calling is still a challenging task due to the complicated relations between different APIs, especially in a context-learning setting without fine-tuning. This paper introduces ``Reverse Chain'', a controllable, target-driven approach designed to empower LLMs with the capability to operate external APIs only via prompts. Recognizing that most LLMs have limited tool-use capabilities, Reverse Chain limits LLMs to executing simple tasks, e.g., API Selection and Argument Completion. Furthermore, to manage a controllable multi-function calling, Reverse Chain adopts a generic rule based on a backward reasoning process. This rule determines when to do API selection or Argument completion. To evaluate the multi-tool-use capability of LLMs, we have released a compositional multi-tool task dataset, available at \url{https://anonymous.4open.science/r/reverse-chain-8681}. Extensive numerical experiments validate the remarkable proficiency of Reverse Chain in managing multiple API calls.
Large Language Models (LLMs) have recently showcased remarkable generalizability in various domains. Despite their extensive knowledge, LLMs still face challenges in efficiently utilizing encoded knowledge to develop accurate and logical reasoning processes. To mitigate this problem, we introduced Hint-before-Solving Prompting (HSP), which guides the model to generate hints (e.g., specific knowledge or key ideas) for solving the problem and then generate solutions containing intermediate reasoning steps. Since HSP is orthogonal to prompting methods (e.g., Chain-of-Thought (CoT)), we applied HSP to CoT, Least-to-Most, Plan-and-Solve, and Standard promptings. The results of extensive experiments on 6 reasoning benchmarks and 4 open-source LLMs demonstrate that HSP can effectively improve the accuracy of reasoning tasks: (1) By applying high-quality hint-enhanced HSP to CoT prompting, Llama2-70B-Chat shows an improvement of 9.7. (2) Beyond exploring training-free LLM capabilities, we built the HSPMATH dataset based on HSP and fine-tuned Llemma-7B, reaching 64.3 accuracy, surpassing GPT-3.5 and WizardMath-13B. We make our code and dataset publicly available at \url{https://github.com/jinlanfu/HSP}.
The integration of Large Language Models (LLMs) into Development Environments (IDEs) has become a focal point in modern software development. LLMs such as OpenAI GPT-3.5/4 and Code Llama offer the potential to significantly augment developer productivity by serving as intelligent, chat-driven programming assistants. However, utilizing LLMs out of the box is unlikely to be optimal for any given scenario. Rather, each system requires the LLM to be honed to its set of heuristics to ensure the best performance. In this paper, we introduce the Copilot evaluation harness: a set of data and tools for evaluating LLM-guided IDE interactions, covering various programming scenarios and languages. We propose our metrics as a more robust and information-dense evaluation than previous state of the art evaluation systems. We design and compute both static and execution based success metrics for scenarios encompassing a wide range of developer tasks, including code generation from natural language (generate), documentation generation from code (doc), test case generation (test), bug-fixing (fix), and workspace understanding and query resolution (workspace). These success metrics are designed to evaluate the performance of LLMs within a given IDE and its respective parameter space. Our learnings from evaluating three common LLMs using these metrics can inform the development and validation of future scenarios in LLM guided IDEs.
Recent advancements in large language models (LLMs) have spurred interest in using them for generating robot programs from natural language, with promising initial results. We investigate the use of LLMs to generate programs for service mobile robots leveraging mobility, perception, and human interaction skills, and where accurate sequencing and ordering of actions is crucial for success. We contribute CodeBotler, an open-source robot-agnostic tool to program service mobile robots from natural language, and RoboEval, a benchmark for evaluating LLMs' capabilities of generating programs to complete service robot tasks. CodeBotler performs program generation via few-shot prompting of LLMs with an embedded domain-specific language (eDSL) in Python, and leverages skill abstractions to deploy generated programs on any general-purpose mobile robot. RoboEval evaluates the correctness of generated programs by checking execution traces starting with multiple initial states, and checking whether the traces satisfy temporal logic properties that encode correctness for each task. RoboEval also includes multiple prompts per task to test for the robustness of program generation. We evaluate several popular state-of-the-art LLMs with the RoboEval benchmark, and perform a thorough analysis of the modes of failures, resulting in a taxonomy that highlights common pitfalls of LLMs at generating robot programs. We release our code and benchmark at https://amrl.cs.utexas.edu/codebotler/.
This work summarizes two ways to accomplish Time-Series (TS) tasks in today's Large Language Model (LLM) context: LLM-for-TS (model-centric) designs and trains a fundamental large model, or fine-tunes a pre-trained LLM for TS data; TS-for-LLM (data-centric) converts TS into a model-friendly representation to enable the pre-trained LLM to handle TS data. Given the lack of data, limited resources, semantic context requirements, and so on, this work focuses on TS-for-LLM, where we aim to activate LLM's ability for TS data by designing a TS embedding method suitable for LLM. The proposed method is named TEST. It first tokenizes TS, builds an encoder to embed TS via instance-wise, feature-wise, and text-prototype-aligned contrast, where the TS embedding space is aligned to LLM embedding layer space, then creates soft prompts to make LLM more open to that embeddings, and finally implements TS tasks using the frozen LLM. We also demonstrate the feasibility of TS-for-LLM through theory and experiments. Experiments are carried out on TS classification, forecasting, and representation tasks using eight frozen LLMs with various structures and sizes. The results show that the pre-trained LLM with TEST strategy can achieve better or comparable performance than today's SOTA TS models and offer benefits for few-shot and generalization. By treating LLM as the pattern machine, TEST can endow LLM's ability to process TS data without compromising language ability. We hope that this study will serve as a foundation for future work to support TS+LLM progress.
In software development, the predominant emphasis on functionality often supersedes security concerns, a trend gaining momentum with AI-driven automation tools like GitHub Copilot. These tools significantly improve developers' efficiency in functional code development. Nevertheless, it remains a notable concern that such tools are also responsible for creating insecure code, predominantly because of pre-training on publicly available repositories with vulnerable code. Moreover, developers are called the "weakest link in the chain" since they have very minimal knowledge of code security. Although existing solutions provide a reasonable solution to vulnerable code, they must adequately describe and educate the developers on code security to ensure that the security issues are not repeated. Therefore we introduce a multipurpose code vulnerability analysis system \texttt{SecRepair}, powered by a large language model, CodeGen2 assisting the developer in identifying and generating fixed code along with a complete description of the vulnerability with a code comment. Our innovative methodology uses a reinforcement learning paradigm to generate code comments augmented by a semantic reward mechanism. Inspired by how humans fix code issues, we propose an instruction-based dataset suitable for vulnerability analysis with LLMs. We further identify zero-day and N-day vulnerabilities in 6 Open Source IoT Operating Systems on GitHub. Our findings underscore that incorporating reinforcement learning coupled with semantic reward augments our model's performance, thereby fortifying its capacity to address code vulnerabilities with improved efficacy.
Data contamination, i.e., the presence of test data from downstream tasks in the training data of large language models (LLMs), is a potential major issue in measuring LLMs' real effectiveness on other tasks. We propose a straightforward yet effective method for identifying data contamination within LLMs. At its core, our approach starts by identifying potential contamination at the instance level; using this information, our approach then assesses wider contamination at the partition level. To estimate contamination of individual instances, we employ "guided instruction:" a prompt consisting of the dataset name, partition type, and the random-length initial segment of a reference instance, asking the LLM to complete it. An instance is flagged as contaminated if the LLM's output either exactly or nearly matches the latter segment of the reference. To understand if an entire partition is contaminated, we propose two ideas. The first idea marks a dataset partition as contaminated if the average overlap score with the reference instances (as measured by ROUGE-L or BLEURT) is statistically significantly better with the completions from guided instruction compared to a "general instruction" that does not include the dataset and partition name. The second idea marks a dataset partition as contaminated if a classifier based on GPT-4 with few-shot in-context learning prompt marks multiple generated completions as exact/near-exact matches of the corresponding reference instances. Our best method achieves an accuracy between 92% and 100% in detecting if an LLM is contaminated with seven datasets, containing train and test/validation partitions, when contrasted with manual evaluation by human experts. Further, our findings indicate that GPT-4 is contaminated with AG News, WNLI, and XSum datasets.
Large-Language-Models (LLMs) are deployed in a wide range of applications, and their response has an increasing social impact. Understanding the non-deliberate(ive) mechanism of LLMs in giving responses is essential in explaining their performance and discerning their biases in real-world applications. This is analogous to human studies, where such inadvertent responses are referred to as sampling. We study this sampling of LLMs in light of value bias and show that the sampling of LLMs tends to favour high-value options. Value bias corresponds to this shift of response from the most likely towards an ideal value represented in the LLM. In fact, this effect can be reproduced even with new entities learnt via in-context prompting. We show that this bias manifests in unexpected places and has implications on relevant application scenarios, like choosing exemplars. The results show that value bias is strong in LLMs across different categories, similar to the results found in human studies.
It has recently been shown that adversarial attacks on large language models (LLMs) can "jailbreak" the model into making harmful statements. In this work, we argue that the spectrum of adversarial attacks on LLMs is much larger than merely jailbreaking. We provide a broad overview of possible attack surfaces and attack goals. Based on a series of concrete examples, we discuss, categorize and systematize attacks that coerce varied unintended behaviors, such as misdirection, model control, denial-of-service, or data extraction. We analyze these attacks in controlled experiments, and find that many of them stem from the practice of pre-training LLMs with coding capabilities, as well as the continued existence of strange "glitch" tokens in common LLM vocabularies that should be removed for security reasons.
The interactive nature of Large Language Models (LLMs) theoretically allows models to refine and improve their answers, yet systematic analysis of the multi-turn behavior of LLMs remains limited. In this paper, we propose the FlipFlop experiment: in the first round of the conversation, an LLM completes a classification task. In a second round, the LLM is challenged with a follow-up phrase like "Are you sure?", offering an opportunity for the model to reflect on its initial answer, and decide whether to confirm or flip its answer. A systematic study of ten LLMs on seven classification tasks reveals that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17% (the FlipFlop effect). We conduct finetuning experiments on an open-source LLM and find that finetuning on synthetically created data can mitigate - reducing performance deterioration by 60% - but not resolve sycophantic behavior entirely. The FlipFlop experiment illustrates the universality of sycophantic behavior in LLMs and provides a robust framework to analyze model behavior and evaluate future models.
Large language models (LLMs) have been successfully adapted for interactive decision-making tasks like web navigation. While achieving decent performance, previous methods implicitly assume a forward-only execution mode for the model, where they only provide oracle trajectories as in-context examples to guide the model on how to reason in the environment. Consequently, the model could not handle more challenging scenarios not covered in the in-context examples, e.g., mistakes, leading to sub-optimal performance. To address this issue, we propose to model the interactive task as state space exploration, where the LLM agent transitions among a pre-defined set of states by performing actions to complete the task. This formulation enables flexible backtracking, allowing the model to recover from errors easily. We evaluate our proposed LLM Agent with State-Space ExploRation (LASER) on both the WebShop task and amazon.com. Experimental results show that LASER significantly outperforms previous methods and closes the gap with human performance on the web navigation task.
In the last years' digitalization process, the creation and management of documents in various domains, particularly in Public Administration (PA), have become increasingly complex and diverse. This complexity arises from the need to handle a wide range of document types, often characterized by semi-structured forms. Semi-structured documents present a fixed set of data without a fixed format. As a consequence, a template-based solution cannot be used, as understanding a document requires the extraction of the data structure. The recent introduction of Large Language Models (LLMs) has enabled the creation of customized text output satisfying user requests. In this work, we propose a novel approach that combines the LLMs with prompt engineering and multi-agent systems for generating new documents compliant with a desired structure. The main contribution of this work concerns replacing the commonly used manual prompting with a task description generated by semantic retrieval from an LLM. The potential of this approach is demonstrated through a series of experiments and case studies, showcasing its effectiveness in real-world PA scenarios.
Large Language Models (LLMs) are trained with a pre-defined context length, restricting their use in scenarios requiring long inputs. Previous efforts for adapting LLMs to a longer length usually requires fine-tuning with this target length (Full-length fine-tuning), suffering intensive training cost. To decouple train length from target length for efficient context window extension, we propose Positional Skip-wisE (PoSE) training that smartly simulates long inputs using a fixed context window. This is achieved by first dividing the original context window into several chunks, then designing distinct skipping bias terms to manipulate the position indices of each chunk. These bias terms and the lengths of each chunk are altered for every training example, allowing the model to adapt to all positions within target length. Experimental results show that PoSE greatly reduces memory and time overhead compared with Full-length fine-tuning, with minimal impact on performance. Leveraging this advantage, we have successfully extended the LLaMA model to 128k tokens using a 2k training context window. Furthermore, we empirically confirm that PoSE is compatible with all RoPE-based LLMs and position interpolation strategies. Notably, our method can potentially support infinite length, limited only by memory usage in inference. With ongoing progress for efficient inference, we believe PoSE can further scale the context window beyond 128k.
Large context window is a desirable feature in large language models (LLMs). However, due to high fine-tuning costs, scarcity of long texts, and catastrophic values introduced by new token positions, current extended context windows are limited to around 128k tokens. This paper introduces LongRoPE that, for the first time, extends the context window of pre-trained LLMs to an impressive 2048k tokens, with up to only 1k fine-tuning steps at within 256k training lengths, while maintaining performance at the original short context window. This is achieved by three key innovations: (i) we identify and exploit two forms of non-uniformities in positional interpolation through an efficient search, providing a better initialization for fine-tuning and enabling an 8x extension in non-fine-tuning scenarios; (ii) we introduce a progressive extension strategy that first fine-tunes a 256k length LLM and then conducts a second positional interpolation on the fine-tuned extended LLM to achieve a 2048k context window; (iii) we readjust LongRoPE on 8k length to recover the short context window performance. Extensive experiments on LLaMA2 and Mistral across various tasks demonstrate the effectiveness of our method. Models extended via LongRoPE retain the original architecture with minor modifications to the positional embedding, and can reuse most pre-existing optimizations.
Unsupervised Text Style Transfer (UTST) has emerged as a critical task within the domain of Natural Language Processing (NLP), aiming to transfer one stylistic aspect of a sentence into another style without changing its semantics, syntax, or other attributes. This task is especially challenging given the intrinsic lack of parallel text pairings. Among existing methods for UTST tasks, attention masking approach and Large Language Models (LLMs) are deemed as two pioneering methods. However, they have shortcomings in generating unsmooth sentences and changing the original contents, respectively. In this paper, we investigate if we can combine these two methods effectively. We propose four ways of interactions, that are pipeline framework with tuned orders; knowledge distillation from LLMs to attention masking model; in-context learning with constructed parallel examples. We empirically show these multi-way interactions can improve the baselines in certain perspective of style strength, content preservation and text fluency. Experiments also demonstrate that simply conducting prompting followed by attention masking-based revision can consistently surpass the other systems, including supervised text style transfer systems. On Yelp-clean and Amazon-clean datasets, it improves the previously best mean metric by 0.5 and 3.0 absolute percentages respectively, and achieves new SOTA results.
Recently, large language models (LLM) based generative AI has been gaining momentum for their impressive high-quality performances in multiple domains, particularly after the release of the ChatGPT. Many believe that they have the potential to perform general-purpose problem-solving in software development and replace human software developers. Nevertheless, there are in a lack of serious investigation into the capability of these LLM techniques in fulfilling software development tasks. In a controlled 2 x 2 between-subject experiment with 109 participants, we examined whether and to what degree working with ChatGPT was helpful in the coding task and typical software development task and how people work with ChatGPT. We found that while ChatGPT performed well in solving simple coding problems, its performance in supporting typical software development tasks was not that good. We also observed the interactions between participants and ChatGPT and found the relations between the interactions and the outcomes. Our study thus provides first-hand insights into using ChatGPT to fulfill software engineering tasks with real-world developers and motivates the need for novel interaction mechanisms that help developers effectively work with large language models to achieve desired outcomes.
The dependence of Natural Language Processing (NLP) intelligent software on Large Language Models (LLMs) is increasingly prominent, underscoring the necessity for robustness testing. Current testing methods focus solely on the robustness of LLM-based software to prompts. Given the complexity and diversity of real-world inputs, studying the robustness of LLMbased software in handling comprehensive inputs (including prompts and examples) is crucial for a thorough understanding of its performance. To this end, this paper introduces RITFIS, a Robust Input Testing Framework for LLM-based Intelligent Software. To our knowledge, RITFIS is the first framework designed to assess the robustness of LLM-based intelligent software against natural language inputs. This framework, based on given threat models and prompts, primarily defines the testing process as a combinatorial optimization problem. Successful test cases are determined by a goal function, creating a transformation space for the original examples through perturbation means, and employing a series of search methods to filter cases that meet both the testing objectives and language constraints. RITFIS, with its modular design, offers a comprehensive method for evaluating the robustness of LLMbased intelligent software. RITFIS adapts 17 automated testing methods, originally designed for Deep Neural Network (DNN)-based intelligent software, to the LLM-based software testing scenario. It demonstrates the effectiveness of RITFIS in evaluating LLM-based intelligent software through empirical validation. However, existing methods generally have limitations, especially when dealing with lengthy texts and structurally complex threat models. Therefore, we conducted a comprehensive analysis based on five metrics and provided insightful testing method optimization strategies, benefiting both researchers and everyday users.
Recent advancements in generative large language models (LLMs) have significantly boosted the performance in natural language processing tasks. However, their efficiency is hampered by the inherent limitations in autoregressive token generation. While parallel decoding with token tree verification, e.g., Medusa, has been proposed to improve decoding parallelism and efficiency, it often struggles with maintaining contextual relationships due to its independent token prediction approach and incurs significant verification overhead, especially with large tree sizes and batch processing. In this paper, we propose ProPD, an efficient LLM parallel decoding framework based on dynamic token tree pruning and generation. ProPD features an advanced early pruning mechanism to efficiently eliminate unpromising token sequences to improve verification efficiency. Additionally, it introduces a dynamic token tree generation algorithm to balance the computation and parallelism of the verification phase in real-time and maximize the overall efficiency across different batch sizes, sequence lengths, and tasks, etc. We verify ProPD across a diverse set of datasets, LLMs, and batch sizes and demonstrate ProPD consistently outperforms existing decoding algorithms by 1.1-3.2x.
Generalizable articulated object manipulation is essential for home-assistant robots. Recent efforts focus on imitation learning from demonstrations or reinforcement learning in simulation, however, due to the prohibitive costs of real-world data collection and precise object simulation, it still remains challenging for these works to achieve broad adaptability across diverse articulated objects. Recently, many works have tried to utilize the strong in-context learning ability of Large Language Models (LLMs) to achieve generalizable robotic manipulation, but most of these researches focus on high-level task planning, sidelining low-level robotic control. In this work, building on the idea that the kinematic structure of the object determines how we can manipulate it, we propose a kinematic-aware prompting framework that prompts LLMs with kinematic knowledge of objects to generate low-level motion trajectory waypoints, supporting various object manipulation. To effectively prompt LLMs with the kinematic structure of different objects, we design a unified kinematic knowledge parser, which represents various articulated objects as a unified textual description containing kinematic joints and contact location. Building upon this unified description, a kinematic-aware planner model is proposed to generate precise 3D manipulation waypoints via a designed kinematic-aware chain-of-thoughts prompting method. Our evaluation spanned 48 instances across 16 distinct categories, revealing that our framework not only outperforms traditional methods on 8 seen categories but also shows a powerful zero-shot capability for 8 unseen articulated object categories. Moreover, the real-world experiments on 7 different object categories prove our framework's adaptability in practical scenarios. Code is released at https://github.com/GeWu-Lab/LLM_articulated_object_manipulation/tree/main.
The copilot framework, which aims to enhance and tailor large language models (LLMs) for specific complex tasks without requiring fine-tuning, is gaining increasing attention from the community. In this paper, we introduce the construction of a Healthcare Copilot designed for medical consultation. The proposed Healthcare Copilot comprises three main components: 1) the Dialogue component, responsible for effective and safe patient interactions; 2) the Memory component, storing both current conversation data and historical patient information; and 3) the Processing component, summarizing the entire dialogue and generating reports. To evaluate the proposed Healthcare Copilot, we implement an auto-evaluation scheme using ChatGPT for two roles: as a virtual patient engaging in dialogue with the copilot, and as an evaluator to assess the quality of the dialogue. Extensive results demonstrate that the proposed Healthcare Copilot significantly enhances the capabilities of general LLMs for medical consultations in terms of inquiry capability, conversational fluency, response accuracy, and safety. Furthermore, we conduct ablation studies to highlight the contribution of each individual module in the Healthcare Copilot. Code will be made publicly available on GitHub.
In this study, we leverage LLM to enhance the semantic analysis and develop similarity metrics for texts, addressing the limitations of traditional unsupervised NLP metrics like ROUGE and BLEU. We develop a framework where LLMs such as GPT-4 are employed for zero-shot text identification and label generation for radiology reports, where the labels are then used as measurements for text similarity. By testing the proposed framework on the MIMIC data, we find that GPT-4 generated labels can significantly improve the semantic similarity assessment, with scores more closely aligned with clinical ground truth than traditional NLP metrics. Our work demonstrates the possibility of conducting semantic analysis of the text data using semi-quantitative reasoning results by the LLMs for highly specialized domains. While the framework is implemented for radiology report similarity analysis, its concept can be extended to other specialized domains as well.
Large language models (LLMs) are effective at answering questions that are clearly asked. However, when faced with ambiguous queries they can act unpredictably and produce incorrect outputs. This underscores the need for the development of intelligent agents capable of asking clarification questions to resolve ambiguities effectively. This capability requires complex understanding, state tracking, reasoning and planning over multiple conversational turns. However, directly measuring this can be challenging. In this paper, we offer a surrogate problem which assesses an LLMs's capability to deduce an entity unknown to itself, but revealed to a judge, by asking the judge a series of queries. This \textit{entity-deducing game} can serve as an evaluation framework to probe the conversational reasoning and planning capabilities of language models. We systematically evaluate various LLMs and discover significant differences in their performance on this task. We find that strong LLMs like GPT-4 outperform human players by a large margin. We further employ Behavior Cloning (BC) to examine whether a weaker model is capable of imitating a stronger model and generalizing to data or domains, using only the demonstrations from a stronger model. We finally propose to use Reinforcement Learning to enhance reasoning and planning capacity of Vicuna models through episodes of game playing, which lead to significant performance improvement. We hope that this problem offers insights into how autonomous agents could be trained to behave more intelligently in ambiguous circumstances.
Large Language Models (LLMs) have demonstrated superior results across a wide range of tasks, and Retrieval-augmented Generation (RAG) is an effective way to enhance the performance by locating relevant information and placing it into the context window of the LLM. However, the relationship between retrievers and LLMs in a RAG is still under-investigated. Most existing work treats the retriever and the LLM as independent components and leaves a gap between retrieving human-"friendly" information and assembling a LLM-"friendly" context. In this work, we examine a novel bridge mechanism. We validate the ranking and selection assumptions of retrievers in the context of RAG and propose a framework that chains together supervised and reinforcement learning to train a bridge model that optimizes the connection between the retriever and the LLM. Empirical results demonstrate the effectiveness of our method in both question-answering and personalized generation tasks.
Large language models (LLMs) offer significant promise as a knowledge source for task learning. Prompt engineering has been shown to be effective for eliciting knowledge from an LLM, but alone it is insufficient for acquiring relevant, situationally grounded knowledge for an embodied agent learning novel tasks. We describe a cognitive-agent approach, STARS, that extends and complements prompt engineering, mitigating its limitations and thus enabling an agent to acquire new task knowledge matched to its native language capabilities, embodiment, environment, and user preferences. The STARS approach is to increase the response space of LLMs and deploy general strategies, embedded within the autonomous agent, to evaluate, repair, and select among candidate responses produced by the LLM. We describe the approach and experiments that show how an agent, by retrieving and evaluating a breadth of responses from the LLM, can achieve 77-94% task completion in one-shot learning without user oversight. The approach achieves 100% task completion when human oversight (such as an indication of preference) is provided. Further, the type of oversight largely shifts from explicit, natural language instruction to simple confirmation/discomfirmation of high-quality responses that have been vetted by the agent before presentation to a user.
The programming skill is one crucial ability for Large Language Models (LLMs), necessitating a deep understanding of programming languages (PLs) and their correlation with natural languages (NLs). We examine the impact of pre-training data on code-focused LLMs' performance by assessing the comment density as a measure of PL-NL alignment. Given the scarcity of code-comment aligned data in pre-training corpora, we introduce a novel data augmentation method that generates comments for existing code, coupled with a data filtering strategy that filters out code data poorly correlated with natural language. We conducted experiments on three code-focused LLMs and observed consistent improvements in performance on two widely-used programming skill benchmarks. Notably, the model trained on the augmented data outperformed both the model used for generating comments and the model further trained on the data without augmentation.
Recently, large language models (LLMs) have demonstrated superior capabilities in understanding and zero-shot learning on textual data, promising significant advances for many text-related domains. In the graph domain, various real-world scenarios also involve textual data, where tasks and node features can be described by text. These text-attributed graphs (TAGs) have broad applications in social media, recommendation systems, etc. Thus, this paper explores how to utilize LLMs to model TAGs. Previous methods for TAG modeling are based on million-scale LMs. When scaled up to billion-scale LLMs, they face huge challenges in computational costs. Additionally, they also ignore the zero-shot inference capabilities of LLMs. Therefore, we propose GraphAdapter, which uses a graph neural network (GNN) as an efficient adapter in collaboration with LLMs to tackle TAGs. In terms of efficiency, the GNN adapter introduces only a few trainable parameters and can be trained with low computation costs. The entire framework is trained using auto-regression on node text (next token prediction). Once trained, GraphAdapter can be seamlessly fine-tuned with task-specific prompts for various downstream tasks. Through extensive experiments across multiple real-world TAGs, GraphAdapter based on Llama 2 gains an average improvement of approximately 5\% in terms of node classification. Furthermore, GraphAdapter can also adapt to other language models, including RoBERTa, GPT-2. The promising results demonstrate that GNNs can serve as effective adapters for LLMs in TAG modeling.
This paper mainly describes a unified system for hallucination detection of LLMs, which wins the second prize in the model-agnostic track of the SemEval-2024 Task 6, and also achieves considerable results in the model-aware track. This task aims to detect hallucination with LLMs for three different text-generation tasks without labeled training data. We utilize prompt engineering and few-shot learning to verify the performance of different LLMs on the validation data. Then we select the LLMs with better performance to generate high-quality weakly supervised training data, which not only satisfies the consistency of different LLMs, but also satisfies the consistency of the optimal LLM with different sampling parameters. Furthermore, we finetune different LLMs by using the constructed training data, and finding that a relatively small LLM can achieve a competitive level of performance in hallucination detection, when compared to the large LLMs and the prompt-based approaches using GPT-4.
Recognizing the imperative to address the reliability and transparency issues of Large Language Models (LLM), this work proposes an LLM maturity model tailored for text-to-query applications. This maturity model seeks to fill the existing void in evaluating LLMs in such applications by incorporating dimensions beyond mere correctness or accuracy. Moreover, this work introduces a real-world use case from the law enforcement domain and showcases QueryIQ, an LLM-powered, domain-specific text-to-query assistant to expedite user workflows and reveal hidden relationship in data.
This study investigates the behaviors of Large Language Models (LLMs) when faced with conflicting prompts versus their internal memory. This will not only help to understand LLMs' decision mechanism but also benefit real-world applications, such as retrieval-augmented generation (RAG). Drawing on cognitive theory, we target the first scenario of decision-making styles where there is no superiority in the conflict and categorize LLMs' preference into dependent, intuitive, and rational/irrational styles. Another scenario of factual robustness considers the correctness of prompt and memory in knowledge-intensive tasks, which can also distinguish if LLMs behave rationally or irrationally in the first scenario. To quantify them, we establish a complete benchmarking framework including a dataset, a robustness evaluation pipeline, and corresponding metrics. Extensive experiments with seven LLMs reveal their varying behaviors. And, with role play intervention, we can change the styles, but different models present distinct adaptivity and upper-bound. One of our key takeaways is to optimize models or the prompts according to the identified style. For instance, RAG models with high role play adaptability may dynamically adjust the interventions according to the quality of retrieval results -- being dependent to better leverage informative context; and, being intuitive when external prompt is noisy.
We investigate the integration of Large Language Models (LLMs) into query encoders to improve dense retrieval without increasing latency and cost, by circumventing the dependency on LLMs at inference time. SoftQE incorporates knowledge from LLMs by mapping embeddings of input queries to those of the LLM-expanded queries. While improvements over various strong baselines on in-domain MS-MARCO metrics are marginal, SoftQE improves performance by 2.83 absolute percentage points on average on five out-of-domain BEIR tasks.
The prediction has served as a crucial scientific method in modern social studies. With the recent advancement of Large Language Models (LLMs), efforts have been made to leverage LLMs to predict the human features in social life, such as presidential voting. These works suggest that LLMs are capable of generating human-like responses. However, we find that the promising performance achieved by previous studies is because of the existence of input shortcut features to the response. In fact, by removing these shortcuts, the performance is reduced dramatically. To further revisit the ability of LLMs, we introduce a novel social prediction task, Soc-PRF Prediction, which utilizes general features as input and simulates real-world social study settings. With the comprehensive investigations on various LLMs, we reveal that LLMs cannot work as expected on social prediction when given general input features without shortcuts. We further investigate possible reasons for this phenomenon that suggest potential ways to enhance LLMs for social prediction.
Introduction: With the rapid advances in large language models (LLMs), there have been numerous new open source as well as commercial models. While recent publications have explored GPT-4 in its application to extracting information of interest from radiology reports, there has not been a real-world comparison of GPT-4 to different leading open-source models. Materials and Methods: Two different and independent datasets were used. The first dataset consists of 540 chest x-ray reports that were created at the Massachusetts General Hospital between July 2019 and July 2021. The second dataset consists of 500 chest x-ray reports from the ImaGenome dataset. We then compared the commercial models GPT-3.5 Turbo and GPT-4 from OpenAI to the open-source models Mistral-7B, Mixtral-8x7B, Llama2-13B, Llama2-70B, QWEN1.5-72B and CheXbert and CheXpert-labeler in their ability to accurately label the presence of multiple findings in x-ray text reports using different prompting techniques. Results: On the ImaGenome dataset, the best performing open-source model was Llama2-70B with micro F1-scores of 0.972 and 0.970 for zero- and few-shot prompts, respectively. GPT-4 achieved micro F1-scores of 0.975 and 0.984, respectively. On the institutional dataset, the best performing open-source model was QWEN1.5-72B with micro F1-scores of 0.952 and 0.965 for zero- and few-shot prompting, respectively. GPT-4 achieved micro F1-scores of 0.975 and 0.973, respectively. Conclusion: In this paper, we show that while GPT-4 is superior to open-source models in zero-shot report labeling, the implementation of few-shot prompting can bring open-source models on par with GPT-4. This shows that open-source models could be a performant and privacy preserving alternative to GPT-4 for the task of radiology report classification.
Fine-tuning large language models can improve task specific performance, although a general understanding of what the fine-tuned model has learned, forgotten and how to trust its predictions is still missing. We derive principled uncertainty quantification for fine-tuned LLMs with posterior approximations using computationally efficient low-rank adaptation ensembles. We analyze three common multiple-choice datasets using low-rank adaptation ensembles based on Mistral-7b, and draw quantitative and qualitative conclusions on their perceived complexity and model efficacy on the different target domains during and after fine-tuning. In particular, backed by the numerical experiments, we hypothesise about signals from entropic uncertainty measures for data domains that are inherently difficult for a given architecture to learn.
CovRL: Fuzzing JavaScript Engines with Coverage-Guided Reinforcement Learning for LLM-based Mutation
Fuzzing is an effective bug-finding technique but it struggles with complex systems like JavaScript engines that demand precise grammatical input. Recently, researchers have adopted language models for context-aware mutation in fuzzing to address this problem. However, existing techniques are limited in utilizing coverage guidance for fuzzing, which is rather performed in a black-box manner. This paper presents a novel technique called CovRL (Coverage-guided Reinforcement Learning) that combines Large Language Models (LLMs) with reinforcement learning from coverage feedback. Our fuzzer, CovRL-Fuzz, integrates coverage feedback directly into the LLM by leveraging the Term Frequency-Inverse Document Frequency (TF-IDF) method to construct a weighted coverage map. This map is key in calculating the fuzzing reward, which is then applied to the LLM-based mutator through reinforcement learning. CovRL-Fuzz, through this approach, enables the generation of test cases that are more likely to discover new coverage areas, thus improving vulnerability detection while minimizing syntax and semantic errors, all without needing extra post-processing. Our evaluation results indicate that CovRL-Fuzz outperforms the state-of-the-art fuzzers in terms of code coverage and bug-finding capabilities: CovRL-Fuzz identified 48 real-world security-related bugs in the latest JavaScript engines, including 39 previously unknown vulnerabilities and 11 CVEs.
Large language models (LLMs) fine-tuned with reinforcement learning from human feedback (RLHF) have been used in some of the most widely deployed AI models to date, such as OpenAI's ChatGPT or Anthropic's Claude. While there has been significant work developing these methods, our understanding of the benefits and downsides of each stage in RLHF is still limited. To fill this gap, we present an extensive analysis of how each stage of the process (i.e. supervised fine-tuning (SFT), reward modelling, and RLHF) affects two key properties: out-of-distribution (OOD) generalisation and output diversity. OOD generalisation is crucial given the wide range of real-world scenarios in which these models are being used, while output diversity refers to the model's ability to generate varied outputs and is important for a variety of use cases. We perform our analysis across two base models on both summarisation and instruction following tasks, the latter being highly relevant for current LLM use cases. We find that RLHF generalises better than SFT to new inputs, particularly as the distribution shift between train and test becomes larger. However, RLHF significantly reduces output diversity compared to SFT across a variety of measures, implying a tradeoff in current LLM fine-tuning methods between generalisation and diversity. Our results provide guidance on which fine-tuning method should be used depending on the application, and show that more research is needed to improve the tradeoff between generalisation and diversity.